HE Lei, ZHAO Lina, YANG Hongwei, MENG Xiang, WANG Ruyue, ZHANG Guifang
Convolutional neural networks (CNNs) have been widely utilized in hyperspectral image (HSI) classification tasks, achieving remarkable performance. However, in existing HSI-CNN methods, the cubic information in HSI is often vectorized, which can compromise the geometric structure of the data. Meanwhile, how to break through the bottleneck of parameter redundancy and immense computation consumption is a hot topic in CNN-based methods. Inspired by these, a stand-alone tensor neural network (SATNN) is proposed, which uses tensor algebra to construct a deep learning framework to replace the convolutional, pooling, and fully connected layers typically found in CNNs. Feature extraction is carried out among the tensor contraction layer (TCOL), tensor pattern product layer (TMPL), tensor replacing the flatten operation, and the fully-connected layer (TRFFC), which can capture the geometric structure and multilinear structure of high-dimensional data. What is important, TCOL can reduce the parameters of LeNet-5 and the hybrid spectral convolution neural network (HybridSN) by 64.46% and 98.24% with little effect on precision. Experiment results on three commonly used hyperspectral imagery datasets demonstrate the effectiveness of HSI-SATNN, with its classification accuracy surpassing that of several CNN-based and tensor-based approaches.
